Target detection and recognition with polarimetric SAR
R.J. Dekker and A.C. van den Broek
TNO Physics and Electronics Laboratory, P.O. Box 96864, 2509 JG The Hague, The Netherlands
ABSTRACT
Target detection and recognition using polarimetric SAR data has been studied by using PHARUS and RAMSES data
collected during the MIMEX campaign. Additionally very high-resolution ISAR data was used. A basic detection and
recognition scheme has been developed, which includes polarimetric speckle-filtering, CFAR detection and the extraction of
geometrical, intensity and polarimetric features. From the SAR images we conclude that polarimetric features can be useful
to discriminate targets from clutter. At resolutions of 1 meter or better, shape and orientation recognition can be obtained
with these features. To classify the targets, other features or other techniques have to be used. Examples are polarimetric
decomposition techniques, of which two have been explored using the ISAR data.
Keywords: SAR, polarimetry, target detection, recognition, polarimetric decompositions
1. INTRODUCTION
The Royal Netherlands Army is currently introducing UAVs carrying electro-optical imaging sensors operating in the
optical and thermal infrared wavelength region. Since also the need for an all-time and all-weather capability has been
identified, interest for SAR imaging systems is growing. In this context a study was defined in which the role of SAR for
ground surveillance has been investigated. In the study presented here we want to evaluate the role of high and medium
resolution millimetre wave or microwave imaging systems for the purpose of military ground surveillance. We have
therefore participated in the MIMEX campaign in order to survey military relocatable targets deployed at the UK
Swynnerton test site. The campaign was prepared and conducted by the NATO research group AC/243 (Panel 3/RSG 20),
predecessor of the present working group on MMW imaging techniques TG14.
With respect to military intelligence we distinguish between wide area surveillance with medium resolution SAR systems
for the purpose of terrain interpretation and target detection, and target acquisition for the purpose of target detection,
classification and recognition. We will focus here mainly on the target classification and recognition. The aspects of terrain
For the study we have used data from the Dutch PHARUS polarimetric
interpretation and detection were discussed
SAR and the French RAMSES high-resolution and polarimetric SAR, collected during the MIMEX campaign. We present
the results of a target classification algorithm for the relocatable targets in the scene. The benefit of polarimetry at even
higher resolutions has been investigated by using US fully polarimetric ISAR data.
The organisation of the paper is as follows. In section 2 we describe the data that was used. In section 3 we introduce a
target detection and recognition scheme, which we applied to the SAR data. In section 4 we discuss polarimetric
decomposition techniques, which we applied to the high-resolution ISAR data. Finally in section 5we give conclusions.
2. DESCRIPTION OF THE DATA
During the MIMEX campaign (30 September to 25 October 1996) images were collected over the Swynnerton test site with
various sensors. Five targets were deployed, including some relocatable rocket-launchers and a ZSU 23-4 tracked antiaircraft vehicle. The Netherlands participated with the fully polarimetric PHARUS system, operating in the C-band. Images
were successfully collected on 24 October 1996 for 3 tracks: two so-called long range (large incidence angles) images with
perpendicular tracks, and one so-called short range image (small incidence angle). The long-range images have been used in
this study because these images showed higher contrast due to longer shadows. Polarimetric calibration (i.e. phase
calibration, cross-talk removal, and gain imbalance calibration) was done by using clutter statistics and antenna
Other author information: R.J.Dekker; Email: [email protected]: Telephone: +31 70 374 0431; Fax: +31 70374 0654. A.C. van den
Broek; Email: [email protected]; Telephone: +31 70 374 0430; Fax: +31 703740654.
178
In Radar Sensor Technology V, Robert Trebits, James L. Kurtz, Editors,
Proceedings of SPIE Vol. 4033 (2000) • 0277-786X/OO/$1 5.00
Downloaded from SPIE Digital Library on 20 Jul 2010 to 195.169.128.3. Terms of Use: http://spiedl.org/terms
measurements.2 Absolute calibration was not performed. France participated with RAMSES (X, Ka and W band) which is
operated by ONERA. It produces higher resolution images in all bands, see Table 1 . Only the X-band system is
polarimetric. The Ka and W band systems are single channel and have incidence angles comparable to that of the PHARUS
system (75°). All images were processed with a maximum number of looks (Table 1). The numbers of looks of the
polarimetric data is given by their polarimetric filtered images (see also section 3. 1). Polarimetric calibration of the X-band
data was done in the same way as the PHARUS data.
Polarimetric ISAR (i.e. inverse SAR) turntable data was provided by ARL. The data consists of six runs at three
wavelengths and two incidence angles of a ZSU 23-4, which was one the relocatable target in the Swynnerton test area, see
Table 2. The data has been processed using a two-dimensional FFT, and cosine squared weighting (tapering 6 dB). We see
that the incidence angle of 78° is comparable to the incidence angles of the C, Ka and W band data (Table 1).
Table 1. SAR data.
Party
System
Band
Frequency
Polarisation
Altitude
Range
Incidence angle
Resolution
Number of looks
TNO-FEL
PHARUS
ONERA
RAMSES
C
X
5.3 0Hz
9.5 GHz
Quad (linear)
2400 m
3500 m
45°
1.Om
3 (PWF)
Quad (linear)
3000 m
12000 m
75°
3.5m
8 (PWF)
Ka
35 0Hz
W
VV
900 m
3500 m
LR
900 m
3500 m
750
750
O.5m
4
1.Om
95 GHz
8
Table 2. ISAR data (ö stands for azimuth resolution).
Party
ARL
Fully pplarimetric stepped fre uency radar
W
Ka
9.25 GHz
34.25 GHz
94.25 GHz
-
System
x
Band
Centre frequency
Bandwidth
Frequency step
Angle sampling interval
Coherence interval (a 10 cm)
Polarisation
Depression angle
Incidence angle
Resolution
Number of looks
i5 1 1 .64 MHz
15 1 1 .64 MHz
15 1 1 .64 MHz
5.928 MHz
5.928 MHz
0.015°
2.4°
Quad (linear)
30° and 12°
60° and 78°
5.928 MHz
0.006°
0.03°
9°
0.9°
0. 1 m
Single Look
3. DETECTION AND RECOGNITION SCHEME
An Automatic Target detection and Recognition (ATR) scheme is generally described by Fig. 2. Initially pre-processing is
performed, which is often speckle filtering in case of SAR data. The second stage involves detection on the filtered image.
Mostly a Constant False-Alarm Rate (CFAR) detector is used. Then the detected pixels are clustered into segments. After
that several features are extracted for each segment, which are used to determine the nature of the segment. This can be to
discriminate between targets and clutter, or between different types of targets. The latter obviously requires features
extracted from higher resolution SAR data (less than 0.5 m). Note that Fig. 1 is a general scheme. Some authors came up
with other schemes like Kreithen et. al.4 They introduced a discriminator before the classifier, to reject natural-clutter false
alarms. Sometimes it is recommended to reduce the number features with a Karhunen-Loeve transform.5
In this paper we have used a straightforward classification procedure to discriminate the relocatable targets from the naturalclutter false alarms. Due to the limited resolution of the SAR data, and the fact that we are dealing with a limited data-set,
179
Downloaded from SPIE Digital Library on 20 Jul 2010 to 195.169.128.3. Terms of Use: http://spiedl.org/terms
we were not able to discriminate between the different types of the targets deployed during MIMEX campaign. Therefore
we evaluated polarimetric decomposition techniques applied to the high-resolution ISAR data. Because TSAR data contains
little clutter, pre-processing, detection and segmentation were omitted here. In the following we discuss each step in the
detection and recognition scheme in more detail.
Pre-processing
Detection
Target Segmentation
[Feature Extraction
+
Classification
I
Figure 1. General ATR system.
3.1. Polarimetric filtering
Detection in SAR imagery is improved when speckle is reduced, since the signal-to-noise ratio increases for the target. We
discriminate between polarimetric and spatial filtering. The first type of filtering has the advantage that the spatial resolution
is not decreased. We therefore focus on the this type of filtering. Several filters were investigated.' The most optimal filter is
the Polarimetric Whitening Filter (PWF). The essence of the filter is the transformation of linear polarimetric data with base
(HH, HV, VV) to a base where the three polarimetric channels are not correlated. The filter produces a weighted average of
these non-correlated channels. The filter for multi-look data is given by:6
y=
(1)
Here : stands for the covariance matrix of the sample to be filtered, and
for the average covariance matrix surrounding
the sample. tr( ) denotes the trace of a matrix. For often a generic covariance matrix is chosen which is representative for
most of the backgrounds of targets. The maximum increase of number of looks that can be obtained with a PWF filter is 3.
3.2. Detection
In the second stage the detection of pixels belonging to a potential target is performed. We have investigated a few CFAR
detectors, in which the value of a sample of the target is compared with that of the average background clutter.' The hollowstencil order-statistics CFAR detector7 was preferred, because it gives better results producing less false alarms than the
other CFAR detectors. It appears that the best results are obtained using the following CFAR equation:
xt —x5o
x80 — x20
> "CFAR
(2)
Here x stands for the sample to be tested and x50 for the median (50% value). x80 - x20 approximates the standard deviation
(80% value minus 20% value). KCFAR stands for the CFAR constant. The stencil is given by Fig. 2.
180
Downloaded from SPIE Digital Library on 20 Jul 2010 to 195.169.128.3. Terms of Use: http://spiedl.org/terms
IIFLITIIII1IIL:
I target
I 1]
clutter
111L
Figure 2. CFAR template with target.
To be able to compare the CFAR results of all images (PHARUS and RAMSES) we used the same CFAR detector, Eq. (2),
with the same CFAR constant KCFAR = 12.5. This value was chosen, since for this constant the false alarm rate due to
speckle is negligible, and since the results on the basis of visual inspection of all images were optimal. The size of the
template was in the order of 30 m for all images, see Table 4.
3.3. Target segmentation
In this stage the pixels that belong to one potential target are clustered. Segmentation is done by iterative merging of pixels
or segments. The merge criterion, to decide if two pixels or segments belong to the same potential target, is given by the
geometrical distance between the pair. The distance used in merging process was in the order of 1 .5m for all images.
3.4. Feature extraction
From all potential target segments a list of features was extracted, see Table 3. The list is subdivided in geometrical,
radiometric and polarimetric features. We evaluated a feature by comparing it with a threshold, which was tuned to the
actual targets in the images, since no independent training data to derive thresholds was available. The most powerful
geometrical feature turned out to be the number of pixels n covering the target. Using this feature only, we were able to
reduce the number of potential target segments with more than 90% (except for PHARUS), see Table 4. We introduced
therefore a two-stage hierarchical classifier, evaluating the number of pixels n in the first stage, and evaluation of the other
features in the second stage.
In the second stage the other geometrical features were analysed. The Minimum Bounding Rectangle (MBR, i.e. the
rectangle which fits closest to the potential target) turned out to be not an effective feature. The ratio between n and the
MBR (i.e. fill rate), and the length of a target segment were more appropriate for discrimination. The length was computed
by averaging the distances between the top-left and bottom-right pixels, and between the top-right and bottom-left pixels, of
a target segment. Its average direction was extracted as additional information, in case the (potential) target turns out to be a
real one. Note that the direction becomes less accurate as the target becomes more horizontal or vertical oriented.
Calculating the orientation by fitting a rectangular around the target, is possibly a better estimate of the direction.
Of the radiometric features, only the mean and the normalised variance (coefficient of variation) were appropriate for
discrimination, except for PHARUS. Apparently the mean intensity for the real targets is significantly higher than that for
false targets like trees. The minimum and maximum, their derived features, and the variance, appeared to be not appropriate
for discrimination, as these features were quite variable for both the relocatable targets as the false alarms. The weightedrank fill ratio4 is a feature that is comparable to the variance. It is calculated by dividing the sum of intensities of the 5%
brightest pixels, by the sum of intensities of all detected pixels. Despite the expectation of robustness it was not helpful here.
Another feature, the fractal dimension4 may perform better, but was not implemented for this study.
Polarimetric features could only be obtained from the fully polarimetric systems (RAMSES X-band and PHARUS C-band).
These features were extracted from the average covariance matrix of the pixels covering a potential target. The powers and
correlation coefficients in these matrices, appeared to be too variable to be discriminative. Also the angle of the co-polarised
correlation coefficient, which is an indicator for the scatter-mechanism (even or odd bounce, see section 4), was too
variable. Inspection of this angle for other man-made structures in the scene like buildings and fences, showed less variable
181
Downloaded from SPIE Digital Library on 20 Jul 2010 to 195.169.128.3. Terms of Use: http://spiedl.org/terms
values for this angle. They were often close to 00 or 1800, in contrast to the values for the relocatable targets. The
variability of this parameter for the targets is probably due to the complexity of the targets in combination with the
resolution: if one resolution cell contains different kinds of scatter mechanisms, the observed angle is an average somewhere
in the interval between 0° and 180°. Man-made structures like fences or buildings, probably show one dominant scattering
mechanism (e.g. fence-ground or wall-ground, both double bounce). Also the power ratios HH-HV and HH-VV were not
helpful for the same reason. When the scattering is due to several scatter mechanisms in one resolution cell the ratios
become less pronounced and closer to 1. Only the power ratios in case of the X-band data turned out to be useful. The
limited resolution of the PHARUS data was a problem using these features.
Table 3 shows interesting information, with respect to the polarisation used. Although the RAMSES W-band system has a
lower resolution than its Ka-band system, it requires less features to discriminate the five relocatable targets form the clutter.
This is possibly caused by its circular polarisation and the higher number of looks. Compared to the linear polarisation VV
of the Ka-band system. the circular polarisation RL (i.e. right helix, left helix) of the W-band system is more sensitive to
odd bounce scattering. This type of scattering is characteristic for flat plates and trihedral corner reflectors, see section 4. A
fully polarimetric system, like the RAMSES X-band and the PI-JARUS system, has the advantage that every possible
receive and transmit polarisation, linear or circular, can be obtained (section 4).
3.5. Classification
As described in the last section a two stage classifier was used. In the first stage the number of pixels n is evaluated and in
the second stage the other features are evaluated. For operational real-time situations this can be interesting because it is
computational quite efficient. Both stages contained simple and straight-forward processes in which the features are directly
compared to a threshold. Some of the thresholds are given in Table 4. Because they were tuned separately for each image
they are different. The tuned values for the thresholds of the area and length look realistic for the relocatable targets, except
in case of the PHARUS data, which area threshold seem too high. No thresholds could be found that fit all images
simultaneously. Secondly, we see that using the RAMSES images, all deployed targets could be found and all false alarms
could be removed. Using PHARUS data some targets are missed and some false alarms stay detected. The only appropriate
feature is the number of pixels used in the first stage. No other features used in its second stage are helpful. The reason for
this is that its resolution of 3.5 m is probably insufficient, whether it is a polarimetric system or not. On the other hand,
initially (after the first stage) the PHARUS images show the least number of false alarms. It must be noted that a sixth target
(a radiometer platform) was deployed on the day PHARUS collected data. No system could make a distinction between
targets. Therefore other techniques have to be studied. One is template matching, comparing the target segments with
templates generated by measurements or signatures from target facet models.
— ___
Figure 3. RAMSES Ka-band classified image (left): 5 targets out of 5: and PHARUS C-band classified image (right): 5 targets out of 6,
I false.
182
Downloaded from SPIE Digital Library on 20 Jul 2010 to 195.169.128.3. Terms of Use: http://spiedl.org/terms
Table 3. List of extracted features. A dot indicates that the feature was used to discriminate the target segments from clutter. A dot
between brackets indicates that the feature could be used but was not necessary (W-band only).
Sensor
Polarisation
Resolution(m)
Geometricalfeatures
n
MBR
nIMBR
Length
Direction
Intensityfeatures
Maximum
Minimum
Mean
Ka
W
VV
0.5
P1
—--—
•
•
1.0
X
C
4p.
1.0
3.5
jp
•
_!_
(•)
(•)
•
•
——
Number of pixels covering the target
Number ofpixels in minimum bounding rectangle
—
Orientation target
On pixels covering target only
•
(•)
•
•
Max/mm
Max/mean
Variance
Variance/mean2
Weighted-rank fill ratio
Normalised, known as coefficient of variation (cv)
Ratio intensity brightest pixels/all pixels
On average covariance matrix of pixels covering target
Polarimetricfeatures
Power HF! (SS*)
-
Power HV (ShvShv*)
Power VV (SS)
Correlation HH-HV (I ShhShVI)
Correlation HV-VV (1 ShSj)
Correlation HH-VV (J ShhS 1)
Angle RH-VV (LSSvv*)
Ratio HH-HV (SS*/ ShvShv*)
Ratio HH-VV (SS*/ SvvSvv*)
-
-
-
-
.
.
Non-zero for rotated dihedral and dipole scattering
Non-zero for rotated dihedral and dipole scattering
1800 for dihedral (man-made) scattering
Normalised
Normalised
•
Table 4. Thresholds of some features, and the number of target segments left after classification stage 1 and 2. In the column of
PHARUS (C-band) both long-range perpendicular images are given.
Sensor
Resolution(m)
Image size (m)
CFARsize(m)
Ka
W
0.5
400
28
1.0
Thresholds
n max area (m2)
26.4
nminarea(m)
Length max (m)
Length mm (m)
500
X
1.0
600
30
30
32.5
12.5
47.3
7.2
90
5
93
143
&8
Sd
594
497
3 14
44
36
5
5
5
5
14
5
5
0
0
0
12.3
C
3.5
720
30
90.0
45.0
-
Numbersoftargets
Potential target detected
Targets classified (stage 1)
Targets classified (stage 2)
Real targets classified
False targets classified
89/56
6/6 After evaluating the number of pixels (n)
After evaluating the other features, see Table 3.
3/5
RAMSES out of 5, PHARUS out of 6
3/1
183
Downloaded from SPIE Digital Library on 20 Jul 2010 to 195.169.128.3. Terms of Use: http://spiedl.org/terms
4. POLARIMETRIC DECOMPOSITIONS
In the previous section we studied straight forward polarimetric features, directly taken form the covariance matrix. In order
to derive features more directly related to the scatter mechanisms, polarimetric decomposition techniques can be used.
Polarimetric decompositions can be divided in stochastic or coherent.4 8-12 Stochastic decompositions are useful when the
average (multi-looked) covariance matrix of a target is observed, but when single-look high-resolution data is available it is
more interesting to use a coherent decomposition. The latter can be applied directly to the scattering matrix (i.e. Sinclair
matrix). A decomposition resolving for the basic scatter mechanisms even and odd bounce, is given.4 Even and odd stand
for the number of times the signal reflects against a surface before it is send back. An example of an even bounce scatterer is
a dihedral reflector. Examples of odd bounce scatterers are a sphere, a flat plate and a trihedral reflector. The decomposition
is best described in the circular basis, which is given by:'2
.
_(SHH—SW) +1 HV
SRR
sP = (SHH
_ SW )
SHy
(3)
SRL _(SHH+Sw)
2
The even/odd decomposition is then given by:4
even =
s( +(s
2
(4)
odd =ISRLI
From the linear point-of-view we see that in case of even bounce scattering HE and VV are 1 800 out of phase. In case of
odd bounce they are in phase. Another coherent decomposition is given by Krogager,"'2 who decomposes the even bounce
component further in diplane (i.e. dihedral) and helix (i.e. circular dipole):
kd =ISUI,kh 1'RR1j'LL1
JSI>JSLLJ
kd=IS,kh=ISLL)—)SRRI
s)>S
(5)
k5 JRLI
Here kd stands for diplane, kh for helix and k for sphere (i.e. flat plate). Note that k is equal to the odd component in Eq. (4).
Helix scattering is not only generated by a circular dipole. Two or more diplane scatterers in one resolution-cell may also
result in helix scattering.'2 Therefore diplane scattering can be regarded as pure double-bounce and helix scattering as a
more complex type of scattering. To compare both decompositions, high-resolution ISAR data of ARL was used. Figure 5
shows a W-band ISAR image (12° depression angle) of the ZSU 23-4, including a photo taken from nearly the same angle.
From the ISAR image series, the following features were used: total energy, percent pure (even, odd, diplane, etc.) and
percent bright.4 The percent pure feature stands for the fraction of all pixels within the target area, for which the main
fraction of energy is even, odd, diplane, helix or sphere. The percent bright feature stands for the fraction of brightest pixels
(i.e. CFAR image) within the target area for which the main fraction of energy is even, odd, diplane, etc. The features were
obtained for the whole range of aspect angles. Some pure and bright TSAR images were selected to view where the
scatterers are located. Figure 5 and 6 show the percent bright even and odd, respectively diplane, sphere and helix of the Wband TSAR series for the 12° depression angle. From both figures we see that the components fluctuate significantly, and
that the helix component is small. We also see that the diplane component is often smaller than the even component,
compared to odd or sphere, due to the helix scattering. Only at some angles diplane and helix are comparable. It is therefore
questionable if the decomposition proposed by Krogager has an advantage in case of percent pure or bright measurements.
Perhaps it is better not to decompose the even bounce component further. Additional research, e.g. studying other features,
184
Downloaded from SPIE Digital Library on 20 Jul 2010 to 195.169.128.3. Terms of Use: http://spiedl.org/terms
will be necessary to confirm or refute this. On the other hand, if we view some pure and bright ISAR images, we see that
helix scattering is often generated by more complicated structures. Some images for instance, show helix scatteringfrom the
tracks and from the cavity where the guns come out of the turret. This supports the idea that helix is a more complex type of
scattering. The Krogager decomposition can be used in combination with classification by template matching, which
possibly gives better results. However, for this purpose it is necessary to study more measurements from more different
targets. Since the effort to obtain such data can be elaborate, ISAR signatures obtained from data generated using a RCS
prediction code in combination with target facet models, can be a solution.
-:;
___________
*
Figure 4. ZSU 23-4 tracked anti-aircraft vehicle on the ARL turntable and corresponding ISAR image (W-band. 12° depression angle).
Figure 5. Percent bright even (solid) and odd (dashed), as a function of the onentation angle (W-band, 12° depression angle).
Figure 6. Percent bright diplane (solid), helix (dot-dashed) and sphere (dashed), as a function of the onentation angle (W-band, 12°
depression angle).
5.
CONCLUSIONS
In this paper we have presented results of our work concerning automatic target detection and recognition with polarimetric
SAR data. A basic detection and recognition scheme was developed, including polarimetric speckle-filtering. CFAR
detection and the extraction of geometrical, intensity and polarimetric features. The SAR data that were usedvaried from
single polarisation linear and circular to fully polarimetric data with a resolution of 0.5 to 3.5 m. For the study of
185
Downloaded from SPIE Digital Library on 20 Jul 2010 to 195.169.128.3. Terms of Use: http://spiedl.org/terms
polarimetric features in particular, we also used high resolution (0.1 m) ISAR data in combination with polarimetric
decompositions.
From the SAR images of a scene containing five relocatable targets, we conclude that next to geometrical and intensity
features, polarimetric features based on the covariance matrix, can be useful to discriminate the targets from clutter. The
resolution must be about 1 m or better. In that case, size and orientation recognition of the target is possible. To discriminate
between targets, other features and other techniques should be used. One potential technique comprises polarimetric features
derived from polarimetric decompositions. Polarimetric decompositions in combination with template matching can be
more efficient. In order to study the robustness of features with respect to the orientation of targets, and to study the
potential to discriminate between targets, more and different-target data-sets are needed. These data-sets should preferably
be polarimetric. One way of obtaining such polarimetric data-sets is to use a RCS prediction code in combination with facet
models.
REFERENCES
1. A.C. van den Broek, R.J. Dekker, A.J.E. Smith, and F.P.P. de Vries, "Target detection with polarimetric C-band SAR",
Proc. Ofthe RTO Sensors and Electronics Technology Panel (SET) Symp., High Resolution Radar Techniques,
Granada, Spain, 22-24 march 1999, RTO-MP-40, AC/323(SET)TP/8, p. 56, 1999.
2. R.J. Dekker and A.C. van den Broek, "Calibration of Polarimetric PHARUS Data", Proc. Ofthe CEOS SAR Workshop,
Noordwijk, The Netherlands, 3-6 February 1998.
3. C.C. Chen and H.C. Andrews, "Multifrequency Imaging of Radar Turntable Data", IEEE Trans. on Aerosp. Electron.
Syst., Vol. 16, No. 1, pp. 15-22, 1980.
4. D.E. Kreithen, S.D. Halversen, and G.J. Owirka, "Discriminating Targets from Clutter", Lincoln Laboratory Journal,
Vol. 6, No. 1, pp. 25-52, 1993.
5. R.O. Duda and P.E. Hart, Pattern Classification and Scene Analysis, John Wiley and Sons, New York, 1973.
6. G. Liu, S. Huang, A. Torre, and F. Rubertone, "The Multilook Polarimetric Whitening Filter (MPWF) for Intensity
Speckle Reduction in Polarimetric SAR Images", IEEE Trans. On Geosci. Remote Sensing, Vol. 36, No. 3, pp. 10161020, 1998.
7. L.M. Novak and S.R. Hesse, "On the performance of order-statistics CFAR detectors", IEEE Conference Record of the
25tI Asilomar Conference on Signals, Systems and Computers, Vol.2, pp. 835-840, 1991.
8. S.R. Cloude and E. Pottier, "A Review of Target Decomposition Theorems in Radar Polarimetry", IEEE Trans. On
Geosci. Remote Sensing, Vol. 34, No. 2, pp. 498-5 1 8, 1996.
9. Y. Dong, B.C. Forster, and C. Ticehurst, "A New Decomposition of Radar Polarization Signatures", IEEE Trans. On
Geosci. Remote Sensing, Vol. 36, No. 3, pp. 933-939, 1998.
10. A. Freeman and S.L. Durden, "A Three-Component Scattering Model for Polarimetric SAR Data", IEEE Trans. On
Geosci. Remote Sensing, Vol. 36, No. 3, pp. 963-973, 1998.
1 1 . E. Krogager, Aspects ofPolarimetric Radar Imaging, Doctoral Thesis, Technical University of Denmark, Copenhagen,
1993.
12. E. Krogager and W.M. Boerner, "On the Importance of Polarimetric Information in Radar Imaging and Classification",
Proc. Of the AGARD Sensor and Propagation Panel Symp., Remote Sensing: A Valuable Source of Information,
Toulouse, France, 22-25 April 1996, AGARD-CP-582, p. 17, 1996.
186
Downloaded from SPIE Digital Library on 20 Jul 2010 to 195.169.128.3. Terms of Use: http://spiedl.org/terms